Today for AI
HOT RADAR
llmHEAT 11.5°

AI CLUSTERED EVENT · 10/9/2026

ByteDance Seed Reveals DeepSeek Long-Context Drift: Chunked KV Cache Compression Causes Phase Sensitivity with 40% Accuracy Fluctuation

1 reports archived1 independent sourcesupdated 10/9/2026, 08:49:58
Synthesis & Latest Updates
1 Sources Cross-Validated

ByteDance Seed team published a paper identifying that DeepSeek's chunked KV cache compression introduces 'phase sensitivity,' causing periodic performance drift in long-context retrieval. The study reveals systematic asymmetry where identical information is easily retrieved at one phase but difficult at another, with accuracy fluctuations reaching up to 40 percentage points in large open-weight models, exposing weaknesses masked by average benchmark scores.

LATEST/ByteDance Seed team published a paper identifying that DeepSeek's chunked KV cache compression introduces 'phase sensitivity,' causing periodic performance drift in long-context retrieval. The study reveals systematic asymmetry where identical information is easily retrieved at one phase but difficult at another, with accuracy fluctuations reaching up to 40 percentage points in large open-weight models, exposing weaknesses masked by average benchmark scores.

TIMELINECoverage timeline

Total 1 reports · Latest first
  1. IT之家 · 智能时代T2·78 pts

    ByteDance Seed Reveals DeepSeek Long-Context Drift: Chunked KV Cache Compression Causes Phase Sensitivity with 40% Accuracy Fluctuation

    Original: 字节 Seed 团队发现 DeepSeek“抽风”原因，长上下文可能性能漂移

    • Chunked KV cache compression reduces memory costs but introduces token phase as a new positional coordinate, leading to periodic retrieval performance fluctuations.
    • DeepSeek-V4 series models exhibit significant phase sensitivity in long-context tasks, with retrieval accuracy differences reaching up to 40% across different phases.
    • Average benchmark scores may mask systematic weaknesses at specific input positions, necessitating attention to fine-grained performance distributions.